How to Measure Anything

Hubbard argues that anything a manager cares about can be measured, and that the objection that something is immeasurable rests on three confusions: about what measurement is, about what the object of measurement is, and about the methods available. He defines measurement as a reduction in uncertainty based on observation, expressed as a probability distribution rather than as a point estimate.

The method has several parts. First, define the decision the measurement is meant to inform, since a measurement that cannot change a decision has no value. Second, express current uncertainty as a calibrated probability interval, and Hubbard supplies calibration training exercises showing that most people are badly overconfident and that calibration is a learnable skill. Third, compute the expected value of information for each variable, which usually shows that the variables people want to measure have low information value and that a few overlooked ones dominate. Fourth, measure only those, using the cheapest method that reduces uncertainty, including small samples, decomposition, controlled experiments and Fermi estimation. He is emphatic that small samples carry far more information than people assume, and gives the rule of five, that the median of any population lies between the smallest and largest of five random samples about ninety-four percent of the time.

Hubbard also attacks common alternatives, particularly weighted scoring models and risk matrices, arguing they add noise rather than information. The book is practical and the calibration and expected-value-of-information material is directly usable. Its weakness is that Hubbard's own consulting method is presented as more settled than the evidence supports, and some case results are asserted rather than documented.